Pith. sign in

REVIEW 4 major objections 4 minor 76 references

Segmenting France Across Four Centuries

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Four centuries of French maps become one aligned, labeled dataset for tracing landscape change.

desk verdict The FRAx4 dataset is a real contribution; the claim that translation+segmentation is the most promising weakly-supervised approach does not survive its own État-Major numbers. read the letter →

arxiv 2505.24824 v1 pith:OGX4O2IW submitted 2025-05-30 cs.CV

Marta López-Rauhut , Hongyu Zhou , Mathieu Aubry , Loic Landrieu This is my paper
classification cs.CV
keywords historicalmapsegmentationlandcoverchangeweaklysupervisedlearningimage-to-imagetranslationCycleGANFrancecartographymulti-centurydatasetU-Net
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces a dataset of historical maps covering all of metropolitan France (548,305 km²) across four centuries, with tiles from the Cassini (18th century), État-Major (19th century), SCAN50 (20th century), and modern IGN collections. It provides modern land-cover labels for every tile and 22,878 km² of manually annotated historical labels for the two oldest collections, making it a benchmark for both fully supervised and weakly supervised segmentation. The authors show that a model trained only on modern labels can already extract meaningful forests, water, roads, and buildings from centuries-old maps, and that translating historical maps into a modern cartographic style first improves this weak-supervision pipeline on the Cassini collection. They then use their best models to reconstruct forest-cover evolution across France from the 18th century to today, aligning with known historical trends such as reforestation and the transformation of the Landes marshlands. If correct, this gives researchers a national-scale, multi-century resource for studying land use and land cover change without requiring expensive annotations for every map.

What carries the argument

The load-bearing object is the dataset itself: four aligned tile sets of 10,952 tiles each covering metropolitan France, sourced from the Cassini maps (1750–1815), the État-Major maps (1820–1866), SCAN50 (1889–1922), and modern vector maps, with modern labels for every tile and 22,878 km² of manually corrected historical labels for the first two collections. Around this dataset, the methods rest on three components: an 8-stage nnU-Net segmentation backbone; a CycleGAN translation module extended with a translation loss that exploits the weak spatial alignment between historical and modern maps; and a dilated IoU metric that tolerates small misalignments when scoring thin features like roads and rivers. The translation-and-segmentation pipeline works by first converting historical maps into a modern cartographic style, then applying a segmentation model trained on modern labels.

What would settle it

Select a fresh set of tiles at random across all of France, annotate them with independent labelers, and compare the supervised and weakly supervised scores on those tiles; if annotator agreement is low or the scores drop sharply relative to the published 470-tile numbers, the benchmark's ground truth is not representative.

Watch

Extended reading notes

Core claim

The central claim is that a unified, national-scale historical map dataset can be assembled from freely available French cartography and used to benchmark segmentation across four centuries. The paper demonstrates that a fully supervised U-Net trained on sparse historical labels achieves the best results (96.7% overall accuracy on Cassini, 91.3% on État-Major), while weakly supervised models that rely only on modern labels are feasible but clearly weaker. Adding a CycleGAN-based image-to-image translation step, together with a weak-alignment translation loss and color- and detail-matched synthetic modern maps, improves the weakly supervised results on the Cassini maps (mean dilated IoU from 26.1 to 36.1). The paper also shows that the resulting predictions can be aggregated into century-scale forest-density maps of France that broadly match historical records.

Load-bearing premise

The whole benchmark evaluation rests on the 470 manually annotated tiles being accurate and representative of the two historical map collections; those labels were assembled from pre-existing annotations and about 160 hours of manual correction, with no inter-annotator agreement reported.

Editorial extensions

If this is right

  • A shared, national-scale benchmark now exists for historical map segmentation, with four centuries of aligned maps, full modern labels, and partial historical labels on 470 tiles.
  • Fully supervised segmentation on sparse historical labels sets a strong reference (96.7% OA on Cassini, 91.3% on État-Major), giving future methods a clear target to beat.
  • Weakly supervised training on modern labels alone is viable, and adding CycleGAN style translation improves Cassini results substantially (mean dIoU 26.1 to 36.1), though not uniformly on État-Major.
  • The best models can produce century-scale forest-cover maps for all of metropolitan France, enabling quantitative reconstruction of reforestation and marshland conversion over roughly 250 years.
  • The two components that drive translation-based gains are adapting modern labels' color and level of detail to each historical collection, and adding the weak-alignment translation loss.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the weak-supervision pipeline generalizes beyond the two annotated collections, it could be applied to SCAN50 or other unlabeled historical map series to produce retrospective land-cover estimates at continental scale, not just for France.
  • Because the 470 evaluation tiles were selected based on pre-existing annotations and terrain diversity, the reported numbers may not represent the full map collections; a random stratified sample of tiles across all terrain types would test this.
  • The dataset's aligned multi-century pairs could serve research beyond segmentation, such as separating true land-cover change from stylistic variation in maps, a problem directly relevant to remote sensing change detection.
  • The absence of reported inter-annotator agreement leaves label noise unquantified; measuring agreement on a subset of the manual labels would sharpen comparisons between the three baselines.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces a large-scale historical map dataset for metropolitan France, comprising aligned tile sets from four sources: 18th-century Cassini maps, 19th-century État-Major maps, 20th-century SCAN50 maps, and modern vector maps, covering 548,305 km2. The authors provide full modern land-cover labels, manual historical labels for 470 tiles (22,878 km2) for Cassini and État-Major, and benchmark three segmentation approaches: fully supervised training on historical labels, direct weakly supervised training on modern labels, and a two-stage CycleGAN-based translation-plus-segmentation pipeline. They report quantitative results on Cassini and État-Major, qualitative results on SCAN50, an ablation of the translation loss and style-matching components, and an application to long-term forest monitoring. The central methodological claim is that translation into a modern cartographic style before segmentation can significantly improve weakly supervised segmentation of historical maps.

Significance. If the dataset is released as described, it is a valuable community resource: it provides an unprecedented national-scale, multi-century, multi-collection benchmark with aligned maps, full modern labels, and a substantial manual historical-label subset. The benchmark is carefully organized, with 7-fold cross-validation for supervised training, explicit exclusion of the 470 evaluation tiles from weakly supervised training, and clear ablations. The quantitative results credibly show that translation helps on Cassini, and the qualitative analyses and forest-monitoring application illustrate plausible downstream use. However, the paper's headline method claim is not supported across both annotated collections, and the État-Major evaluation has a known label-incompleteness problem that undermines the numerical comparison for that collection.

major comments (4)
  1. [§5.2, Table 2] The État-Major road labels are incomplete, as the paper itself states: 'the annotations for these maps only include main roads, predicting secondary roads is penalized even when they are correct.' This directly invalidates the road dIoU comparison (direct weakly-supervised 6.3 vs. translation+segmentation 5.3) and the mean dIoU comparison (38.6 vs. 28.7) for État-Major. Consequently, the conclusion that translation+segmentation is 'arguably the most promising weakly-supervised approach' is not supported by the quantitative evaluation for État-Major, and the reported numbers cannot be treated as a fair characterization of method performance. The authors should either correct the État-Major road annotations, exclude roads from the État-Major quantitative comparison, or clearly restrict the quantitative claim of translation superiority to Cassini.
  2. [Abstract and §6] The abstract and conclusion assert that translating historical maps to a modern style 'can significantly enhance segmentation performance,' but Table 2 shows the opposite for État-Major: translation+segmentation has lower OA (78.4 vs. 83.3) and lower mean dIoU (28.7 vs. 38.6) than direct weakly supervised segmentation. This is an internal inconsistency between the stated claim and the reported results. The claim needs to be qualified to say that the enhancement is observed on Cassini and that on État-Major translation+segmentation is qualitatively sharper but numerically weaker, or else the conclusion must be revised to match the evidence.
  3. [§3, §5.1] The 470 evaluation tiles were selected 'based on the extent of pre-existing annotations and terrain diversity' and no inter-annotator agreement or representativeness analysis is reported. Because these tiles are the sole quantitative ground truth for the benchmark, selection bias could make the reported dIoU and OA unrepresentative of the full Cassini and État-Major collections. The authors should report the distribution of terrain types, geographic coverage, and annotation quality metrics for the evaluation tiles, or otherwise justify that the selection does not compromise the benchmark's conclusions.
  4. [§5.1, Table 2] The paper reports a single aggregated performance number for each baseline, with only a sentence noting standard deviations (below 1.3 for Cassini and 4.3 for État-Major). The État-Major differences between direct weakly-supervised and translation+segmentation are not accompanied by significance tests or per-fold/per-run results, so it is unclear whether the observed numerical gaps are reliable. I request confidence intervals, per-fold breakdowns, or per-run results for at least the headline comparisons.
minor comments (4)
  1. [Eq. (1), §4.3] The translation-loss formula writes expectations over x∼X and y∼Y but uses the paired sample (x,y) inside the norm; the pairing (i.e., aligned historical and modern tiles) should be made explicit in the notation and in the definition of the expectation.
  2. [Figure 1] The caption says the 21st-century map is 'vectorized,' but the figure appears to show a raster rendering; clarify whether the source is a vector map rendered to a raster tile for alignment.
  3. [Table 1] The 'Ours' row uses checkmark and percentage symbols that are not self-explanatory; adding a legend or footnote for the symbols would improve readability.
  4. [§5.2] The statement that translation+segmentation 'delineates individual buildings, forest boundaries, roads, and rivers more accurately' for État-Major is a qualitative judgment presented without a quantitative counterpart; if retained, it should be supported by a focused qualitative evaluation protocol or examples with consistent zoom and overlay.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the benchmarked models are evaluated on external historical labels not used to train the weakly supervised baselines, and no load-bearing derivation reduces to a fitted input or self-citation.

full rationale

The derivation chain is self-contained. The supervised baseline is trained on the 470 manually annotated historical tiles under 7-fold cross-validation, so held-out predictions are not fitted to the evaluation tiles. The weakly supervised baselines use only modern vector labels and historical map images; the historical label tiles are explicitly excluded from their training: "We exclude the 470 manually annotated tiles used for evaluation and split the remaining 10,482 historical tiles into 9,096 training tiles and 1,386 validation tiles." The CycleGAN translation module is trained on unaligned image pairs with fixed hyperparameters (λcyc = 1.0, λid = 0.5, λtran = 0.5) and no fitting to historical ground truth; the additional loss in Eq. (1) uses only image-domain distances, not label agreement. The paper contains no load-bearing self-citation and invokes no uniqueness theorem from the authors' prior work. The known weakness that the État-Major ground truth omits secondary roads ("the annotations for these maps only include main roads, predicting secondary roads is penalized even when they are correct") undermines the validity of some benchmark comparisons, but that is a measurement-validity concern rather than circularity: the reported evaluations are not constructed from the predictions or from parameters fitted to the target labels.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on standard deep learning assumptions plus three domain assumptions: label accuracy, geometric alignment, and the usefulness of modern labels as weak supervision. The only hand-set numbers are loss weights, dIoU margin, and crop sizes, none of which are fitted to the evaluation data. No new physical or conceptual entities are introduced.

free parameters (3)
  • CycleGAN loss weights = lambda_cyc=1.0, lambda_id=0.5, lambda_tran=0.5
    Hand-set hyperparameters for the translation + segmentation baseline (Section 4.3). The ablation shows Ltran helps on Cassini but is neutral or slightly negative on État-Major.
  • dIoU dilation margin w = 3 pixels
    Hand-chosen tolerance for evaluation metric (Section 5.1). This affects reported dIoU values but is applied uniformly to all baselines.
  • U-Net crop sizes = 1000x1000 (Cassini), 500x500 (État-Major/SCAN50)
    Chosen to match map scale (Section 4.3).
assumptions (4)
  • domain assumption The manual historical labels for Cassini and État-Major are accurate enough to serve as ground truth.
    Labels are taken from GeoHistoricalData and IGN and manually corrected over 160 hours, but no inter-annotator agreement or independent validation is reported. Section 3, Historical Labels.
  • domain assumption The four map collections can be geometrically aligned with modern maps sufficiently for pixel-wise supervision and dIoU evaluation.
    The paper states Cassini has near-perfect alignment; alignment for other maps is implied by the common georeferenced coordinate system. Section 3, Composition.
  • domain assumption Modern land cover labels provide a useful supervisory signal for historical segmentation despite substantial temporal changes.
    This is stated as a hypothesis in Section 4.2: 'we hypothesize that certain permanent landscape features... provide a supervisory signal.' It is empirically tested but not guaranteed.
  • domain assumption The selected 470 tiles for historical labels are representative of France's territory and map diversity.
    Tiles were chosen 'based on the extent of pre-existing annotations and terrain diversity' (Section 3), which may introduce selection bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Segmenting France Across Four Centuries." pith.science (2026). https://pith.science/paper/OGX4O2IW

@misc{pith2026250524824,
  author       = {Pith},
  title        = {Pith review of: Segmenting France Across Four Centuries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OGX4O2IW}},
  note         = {Machine review of arXiv:2505.24824}
}
read the original abstract

Historical maps offer an invaluable perspective into territory evolution across past centuries--long before satellite or remote sensing technologies existed. Deep learning methods have shown promising results in segmenting historical maps, but publicly available datasets typically focus on a single map type or period, require extensive and costly annotations, and are not suited for nationwide, long-term analyses. In this paper, we introduce a new dataset of historical maps tailored for analyzing large-scale, long-term land use and land cover evolution with limited annotations. Spanning metropolitan France (548,305 km^2), our dataset contains three map collections from the 18th, 19th, and 20th centuries. We provide both comprehensive modern labels and 22,878 km^2 of manually annotated historical labels for the 18th and 19th century maps. Our dataset illustrates the complexity of the segmentation task, featuring stylistic inconsistencies, interpretive ambiguities, and significant landscape changes (e.g., marshlands disappearing in favor of forests). We assess the difficulty of these challenges by benchmarking three approaches: a fully-supervised model trained with historical labels, and two weakly-supervised models that rely only on modern annotations. The latter either use the modern labels directly or first perform image-to-image translation to address the stylistic gap between historical and contemporary maps. Finally, we discuss how these methods can support long-term environment monitoring, offering insights into centuries of landscape transformation. Our official project repository is publicly available at https://github.com/Archiel19/FRAx4.git.

Figures

Figures reproduced from arXiv: 2505.24824 by the authors.

Figure 1
Figure 1. Semantic Segmentation of Historical Maps of France Spanning Four Centuries. Top row: aligned input maps of the same area. The 21st - century map is a vectorized present-day map. Bottom row: semantic segmenta￾tion of each input map into four classes. In this work, we introduce a large-scale dataset for historical map analy￾sis, covering four centuries of cartography over metropolitan France, as shown in fig. 1. Our d… view at source ↗
Figure 2
Figure 2. Historical Maps and Labels. Our dataset covers three collections of historical maps. We provide manual annotations of the highlighted regions of France for the Cassini and État-Major collections. 4 Method We introduce three baseline approaches for segmenting historical maps, illus￾trated in fig. 3. Our first approach is a fully-supervised model trained using scarce historical annotations (sec. 4.1). We also propose … view at source ↗
Figure 3
Figure 3. Historical Map Segmentation Baselines. We train three semantic segmentation baselines to predict historical map labels, one supervised (fig. 3a) and two weakly-supervised (fig. 3b and fig. 3c). 4.2 Weakly-Supervised Segmentation Manually annotating historical maps is labor-intensive, making it impractical for large-scale studies. In contrast, modern land cover labels can be freely obtained from open access vector-ba… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Modern Maps and Labels Workflow. We tailor the level of detail and appearance of the modern vector map labels to each historical map collection, and then derive a synthetic modern map by colorizing the labels. lenging when the temporal gap between historical maps and m…
Figure 5
Figure 5. Figure 5: Qualitative Semantic Segmentation Results. Predicted segmenta￾tion of our three baselines and SAM [34], on the same area across three centuries. Evaluation Setting. To segment large tiles, we split them into patches of the same size used during training, without overla…
Figure 6
Figure 6. Figure 6: Image-to-Image Translation. Examples of historical-to-modern and modern-to-historical generation with our image translation model (sec. 4.3). accurately. Because the annotations for these maps only include main roads, predicting secondary roads is penalized even when t…
Figure 7
Figure 7. Figure 7: Qualitative Ablation Study. Impact of color and level of detail match￾ing of the modern maps (“Match”, see fig. 4), and translation loss (Ltran) on segmentation quality for our translation + segmentation baseline. 18th century to the present day. For each baseline and …
Figure 8
Figure 8. Figure 8: Predicted Forest Cover over France Across Four Centuries. The framed map at the bottom right is the current forest density derived from con￾temporary sources. The colored ovals highlight key areas in our analysis: the Landes ○, Pyrenees ○, Massif Central ○, and Provenc…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

76 extracted references · 71 canonical work pages

  1. [1]

    IEEE Geoscience and Remote Sensing Letters 19, 1–4 (2022)

    Andrade, H.J.A., Fernandes, B.J.T.: Synthesis of satellite-like urban images from historical maps using conditional GAN. IEEE Geoscience and Remote Sensing Letters 19, 1–4 (2022)

    work page 2022

  2. [2]

    In: Digital Humanities Conference (2019)

    Ares Oliveira, S., Di Lenardo, I., Tourenc, B., Kaplan, F.: A deep learning approach to cadastral computing. In: Digital Humanities Conference (2019)

    work page 2019

  3. [3]

    In: Premiere Annual Conference of the International Alliance of Digital Humanities Organizations (2017)

    Ares Oliveira, S., di Lenardo, I., Kaplan, F.: Machine vision algorithms on cadaster plans. In: Premiere Annual Conference of the International Alliance of Digital Humanities Organizations (2017)

    work page 2017

  4. [4]

    Abstracts of the ICA7, 7 (2024)

    Arzoumanidis, L., Fethers, J.O., Mudiyanselage, S.H., Dehbi, Y.: Deep generation of synthetic training data for the automated extraction of semantic knowledge from historical maps. Abstracts of the ICA7, 7 (2024)

    work page 2024

  5. [5]

    IEEE Geoscience and Remote Sensing Letters19, 1–5 (2022)

    Avcı, C., Sertel, E., Kabadayı, M.E.: Deep learning-based road extraction from historical maps. IEEE Geoscience and Remote Sensing Letters19, 1–5 (2022)

    work page 2022

  6. [6]

    In: IGARSS 2024- 2024 IEEE International Geoscience and Remote Sensing Symposium

    Briand, S., Weissgerber, F., Fournier, P., Weissgerber, M.: LULC segmentation in historical images under domain shift: An empirical study. In: IGARSS 2024- 2024 IEEE International Geoscience and Remote Sensing Symposium. pp. 488–492 (2024)

    work page 2024

  7. [7]

    IEEE Access9, 62847–62856 (2021)

    Can, Y.S., Gerrits, P.J., Kabadayi, M.E.: Automatic detection of road types from the third military mapping survey of Austria-Hungary historical map series with deep convolutional neural networks. IEEE Access9, 62847–62856 (2021)

    work page 2021

  8. [8]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Cheng, B., Girshick, R., Dollár, P., Berg, A.C., Kirillov, A.: Boundary IoU: Improving object-centric image segmentation evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15334– 15342 (2021)

    work page 2021

Show all 76 references
  1. [9]

    Chiang,Y.Y.,Duan,W.,Leyk,S.,Uhl,J.H.,Knoblock,C.A.:TrainingDeepLearn- ing Models for Geographic Feature Recognition from Historical Maps, pp. 65–98. Springer (2020)

    2020

  2. [10]

    In: Graphics Recognition

    Chiang, Y.Y., Leyk, S., Knoblock, C.A.: Efficient and robust graphics recognition from historical maps. In: Graphics Recognition. New Trends and Challenges. pp. 25–35 (2013)

    2013

  3. [11]

    ACM Computing Surveys47(1), 1–44 (2014) Segmenting France Across Four Centuries 17

    Chiang, Y.Y., Leyk, S., Knoblock, C.A.: A survey of digital map processing tech- niques. ACM Computing Surveys47(1), 1–44 (2014) Segmenting France Across Four Centuries 17

    2014

  4. [12]

    International Journal of Cartography8(1), 18–36 (2022)

    Christophe, S., Mermet, S., Laurent, M., Touya, G.: Neural map style transfer exploration with GANs. International Journal of Cartography8(1), 18–36 (2022)

    2022

  5. [13]

    Revue forestière française48(6), 547–562 (1996)

    Cinotti, B.: Evolution des surfaces boisées en France: Proposition de reconstitution depuis le début du XIXe siècle. Revue forestière française48(6), 547–562 (1996)

    1996

  6. [14]

    Duan, W., Chiang, Y.Y., Leyk, S., Uhl, J.H., Knoblock, C.A.: Automatic alignment of contemporary vector data and georeferenced historical maps using reinforcement learning.InternationalJournalofGeographical InformationScience 34(4),824–849 (2020)

    2020

  7. [15]

    ISPRS International Journal of Geo-Information10(8), 492 (2021)

    Ekim, B., Sertel, E., Kabadayı, M.E.: Automatic road extraction from historical maps using deep learning techniques: A regional case study of Turkey in a German World War II map. ISPRS International Journal of Geo-Information10(8), 492 (2021)

    2021

  8. [16]

    Scientific Reports12(1), 9077 (2022)

    Ettehadi Osgouei, P., Sertel, E., Kabadayı, M.E.: Integrated usage of histori- cal geospatial data and modern satellite images reveal long-term land use/cover changes in Bursa/Turkey, 1858–2020. Scientific Reports12(1), 9077 (2022)

    2022

  9. [17]

    empty space

    Ford, C.: The environmental transformation of “empty space”: From desert to forest in the Landes of southwestern France. Comparative Studies in Society and History 65(2), 422–445 (2023)

    2023

  10. [18]

    Applied Geography 59, 43–55 (2015)

    Fuchs, R., Verburg, P.H., Clevers, J.G., Herold, M.: The potential of old maps and encyclopaedias for reconstructing historic European land cover/use change. Applied Geography 59, 43–55 (2015)

    2015

  11. [19]

    Archaeological Prospection28(2), 187–199 (2021)

    Garcia-Molsosa, A., Orengo, H.A., Lawrence, D., Philip, G., Hopper, K., Petrie, C.A.: Potential of deep learning segmentation for the extraction of archaeologi- cal features from historical map series. Archaeological Prospection28(2), 187–199 (2021)

    2021

  12. [20]

    https: //geohistoricaldata.org, accessed: 2024-11-4

    GeoHistoricalData Project: GeoHistoricalData project website. https: //geohistoricaldata.org, accessed: 2024-11-4

    2024

  13. [21]

    International Journal of Remote Sensing40(9), 3647–3672 (2019)

    Gong, M., Niu, X., Zhan, T., Zhang, M.: A coupling translation network for change detection in heterogeneous images. International Journal of Remote Sensing40(9), 3647–3672 (2019)

    2019

  14. [22]

    Communications of the ACM 63(11), 139–144 (2020)

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020)

    2020

  15. [23]

    ISPRS International Journal of Geo-Information8(1), 2 (2018)

    Hecht, R., Herold, H., Behnisch, M., Jehling, M.: Mapping long-term dynamics of population and dwellings based on a multi-temporal analysis of urban morpholo- gies. ISPRS International Journal of Geo-Information8(1), 2 (2018)

    2018

  16. [24]

    Transactions in GIS24(2), 442–461 (2020)

    Heitzler, M., Hurni, L.: Cartographic reconstruction of building footprints from historical maps: A study on the Swiss Siegfried map. Transactions in GIS24(2), 442–461 (2020)

    2020

  17. [25]

    Geographic Information Science at the Heart of Europe pp

    Herrault, P.A., Sheeren, D., Fauvel, M., Paegelow, M.: Automatic extraction of forests from historical maps based on unsupervised classification in the CIELab color space. Geographic Information Science at the Heart of Europe pp. 95–112 (2013)

    2013

  18. [26]

    In: Proceedings of the 6th ACM SIGSPATIAL International Workshop on Geospatial Humanities

    Hosseini, K., Wilson, D.C., Beelen, K., McDonough, K.: MapReader: A computer vision pipeline for the semantic exploration of maps at scale. In: Proceedings of the 6th ACM SIGSPATIAL International Workshop on Geospatial Humanities. pp. 8–19 (2022)

    2022

  19. [27]

    IGN (1948)

    Huguenin, M.: Historique de la cartographie de la nouvelle carte de France. IGN (1948)

    1948

  20. [28]

    https://geoservices.ign.fr/scanhisto, accessed: 2025-3-4 18 M

    IGN: SCAN Historique. https://geoservices.ign.fr/scanhisto, accessed: 2025-3-4 18 M. López-Rauhut et al

    2025

  21. [29]

    Nature Methods 18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: A self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203–211 (2021)

    2021

  22. [30]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with condi- tional adversarial networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1125–1134 (2017)

    2017

  23. [31]

    In: Proceedings of Automatic Vectorisation of Historical Maps

    Jiao, C., Heitzler, M., Hurni, L.: Extracting wetlands from Swiss historical maps with convolutional neural networks. In: Proceedings of Automatic Vectorisation of Historical Maps. pp. 33–38 (2020)

    2020

  24. [32]

    International Journal of Applied Earth Observation and Geoinformation 113, 102980 (2022)

    Jiao, C., Heitzler, M., Hurni, L.: A fast and effective deep learning approach for road extraction from historical maps by automatically generating training data with symbol reconstruction. International Journal of Applied Earth Observation and Geoinformation 113, 102980 (2022)

    2022

  25. [33]

    IEEE Transactions on Pattern Analysis and Machine Intelligence25(1), 18–31 (2003)

    Khotanzad, A., Zink, E.: Contour line and geographic feature extraction from USGS color topographical paper maps. IEEE Transactions on Pattern Analysis and Machine Intelligence25(1), 18–31 (2003)

    2003

  26. [34]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015–4026 (2023)

    2023

  27. [35]

    Computers & Graphics 35(2), 242–249 (2011)

    Laycock, S.D., Brown, P.G., Laycock, R.G., Day, A.M.: Aligning archive maps and extracting footprints for analysis of historic urban environments. Computers & Graphics 35(2), 242–249 (2011)

    2011

  28. [36]

    In: Graphics Recognition

    Leyk, S.: Segmentation of colour layers in historical maps based on hierarchical colour sampling. In: Graphics Recognition. Achievements, Challenges, and Evolu- tion. pp. 231–241 (2010)

    2010

  29. [37]

    Sensors20(11), 3119 (2020)

    Li, J., Chen, Z., Zhao, X., Shao, L.: MapGAN: An intelligent generation model for network tile maps. Sensors20(11), 3119 (2020)

    2020

  30. [38]

    In: Proceed- ings of the 4th ACM SIGSPATIAL International Workshop on AI for Geographic Knowledge Discovery

    Li, Z., Guan, R., Yu, Q., Chiang, Y.Y., Knoblock, C.A.: Synthetic map generation to provide unlimited training data for historical map text detection. In: Proceed- ings of the 4th ACM SIGSPATIAL International Workshop on AI for Geographic Knowledge Discovery. p. 17–26 (2021)

    2021

  31. [39]

    Ecological Indicators 95, 595–605 (2018)

    Liu, D., Toman, E., Fuller, Z., Chen, G., Londo, A., Zhang, X., Zhao, K.: Integra- tion of historical map and aerial imagery to characterize long-term land-use change and landscape dynamics: An object-based analysis via random forests. Ecological Indicators 95, 595–605 (2018)

    2018

  32. [40]

    Remote Sensing 12(20), 3421 (2020)

    Liu, T., Miao, Q., Xu, P., Zhang, S.: Superpixel-based shallow convolutional neural network (SSCNN) for scanned topographic map segmentation. Remote Sensing 12(20), 3421 (2020)

    2020

  33. [41]

    IEEE Transactions on Geoscience and Remote Sensing 60, 1–22 (2021)

    Luppino, L.T., Kampffmeyer, M., Bianchi, F.M., Moser, G., Serpico, S.B., Jenssen, R., Anfinsen, S.N.: Deep image translation with an affinity-based change prior for unsupervised multimodal change detection. IEEE Transactions on Geoscience and Remote Sensing 60, 1–22 (2021)

    2021

  34. [42]

    Journal of Maps19(1), 2225071 (2023)

    Martinez,T.,Hammoumi,A.,Ducret,G.,Moreaud,M.,Deschamps,R.,Piegay,H., Berger, J.F.: Deep learning ancient map segmentation to assess historical landscape changes. Journal of Maps19(1), 2225071 (2023)

    2023

  35. [43]

    Cambridge University Press (2015)

    Matteson, K.: Forests in revolutionary France: Conservation, community, and con- flict, 1669–1848. Cambridge University Press (2015)

    2015

  36. [44]

    Remote Sensing 12(24), 4145 (2020)

    Maxwell, A.E., Bester, M.S., Guillen, L.A., Ramezan, C.A., Carpinello, D.J., Fan, Y., Hartley, F.M., Maynard, S.M., Pyron, J.L.: Semantic segmentation deep learn- Segmenting France Across Four Centuries 19 ing for extracting surface mine extents from historic topographic maps....

    2020

  37. [45]

    Ambio 52(11), 1777–1792 (2023)

    Mäyrä, J., Kivinen, S., Keski-Saari, S., Poikolainen, L., Kumpula, T.: Utilizing historical maps in identification of long-term land use and land cover changes. Ambio 52(11), 1777–1792 (2023)

    2023

  38. [46]

    IEEE Transactions on Geoscience and Remote Sensing 54(11), 6265–6279 (2016)

    Miao, Q., Liu, T., Song, J., Gong, M., Yang, Y.: Guided superpixel method for to- pographic map processing. IEEE Transactions on Geoscience and Remote Sensing 54(11), 6265–6279 (2016)

    2016

  39. [47]

    Mühlematter, D.J., Schweizer, S., Jiao, C., Xia, X., Heitzler, M., Hurni, L.: Proba- bilistic road classification in historical maps using synthetic data and deep learning (2024), arXiv:2410.02250

    2024 arXiv

  40. [48]

    IEEE Geoscience and Remote Sensing Letters 16(1), 45–49 (2018)

    Niu, X., Gong, M., Zhan, T., Yang, Y.: A conditional adversarial network for change detection in heterogeneous images. IEEE Geoscience and Remote Sensing Letters 16(1), 45–49 (2018)

    2018

  41. [49]

    Ecological Indicators 158, 111363 (2024)

    O’Hara, R., Marwaha, R., Zimmerman, J., Saunders, M., Green, S.: Unleashing the power of old maps: Extracting symbology from nineteenth century maps using convolutional neural networks to quantify modern land use on historic wetlands. Ecological Indicators 158, 111363 (2024)

    2024

  42. [50]

    Scientific Data 2(1), 1–7 (2015)

    Perret, J., Gribaudi, M., Barthelemy, M.: Roads and cities of 18th century France. Scientific Data 2(1), 1–7 (2015)

    2015

  43. [51]

    In: CEUR Workshop Proceedings

    Petitpierre, R.G., Kaplan, F., Di Lenardo, I.: Generic semantic segmentation of historical maps. In: CEUR Workshop Proceedings. pp. 228–248 (2021)

    2021

  44. [52]

    Ecological Engineering139, 105580 (2019)

    Picuno, P., Cillis, G., Statuto, D.: Investigating the time evolution of a rural land- scape:Howhistoricalmapsmayprovideenvironmentalinformationwhenprocessed using a GIS. Ecological Engineering139, 105580 (2019)

    2019

  45. [53]

    Cartography and Geographic Information Science43(3), 250–265 (2016)

    Pindozzi, S., Cervelli, E., Capolupo, A., Okello, C., Boccia, L.: Using historical maps to analyze two hundred years of land cover changes: Case study of Sorrento peninsula (south Italy). Cartography and Geographic Information Science43(3), 250–265 (2016)

    2016

  46. [54]

    In: Medical Image Computing and Computer-Assisted Intervention

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention. pp. 234–241 (2015)

    2015

  47. [55]

    International Journal of Geographical Information Science34(5), 947–968 (2020)

    Saeedimoghaddam, M., Stepinski, T.F.: Automatic extraction of road intersection points from USGS historical map series using deep convolutional neural networks. International Journal of Geographical Information Science34(5), 947–968 (2020)

    2020

  48. [56]

    Computers, Environment and Urban Systems44, 47–58 (2014)

    San-Antonio-Gómez, C., Velilla, C., Manzano-Agugliaro, F.: Urban and landscape changes through historical maps: The Real Sitio of Aranjuez (1775–2005), a case study. Computers, Environment and Urban Systems44, 47–58 (2014)

    2014

  49. [57]

    The North American Review 212(778), 378–386 (1920)

    Scheifley, W.H.: The depleted forests of France. The North American Review 212(778), 378–386 (1920)

    1920

  50. [58]

    Serra, J.: Image analysis and mathematical morphology (1983)

    1983

  51. [59]

    IEEE Transactions on Geoscience and Remote Sensing61, 1–19 (2023)

    Song, J., Chen, H., Du, C., Li, J.: Semi-MapGen: Translation of remote sensing image into map via semisupervised adversarial learning. IEEE Transactions on Geoscience and Remote Sensing61, 1–19 (2023)

    2023

  52. [60]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing14, 2341–2357 (2021)

    Song, J., Li, J., Chen, H., Wu, J.: MapGen-GAN: A fast translator for remote sens- ing image to map via unsupervised adversarial learning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing14, 2341–2357 (2021)

    2021

  53. [61]

    Remote Sensing14(4), 919 (2022) 20 M

    Song, J., Li, J., Chen, H., Wu, J.: RSMT: A remote sensing image-to-map trans- lation model via adversarial deep transfer learning. Remote Sensing14(4), 919 (2022) 20 M. López-Rauhut et al

    2022

  54. [62]

    Ecological Informatics68, 101557 (2022)

    Ståhl, N., Weimann, L.: Identifying wetland areas in historical maps using deep convolutional neural networks. Ecological Informatics68, 101557 (2022)

    2022

  55. [63]

    Remote Sensing 16(19), 3716 (2024)

    Tian, J., Wu, J., Chen, H., Ma, M.: MapGen-Diff: An end-to-end remote sensing image to map generator via denoising diffusion bridge model. Remote Sensing 16(19), 3716 (2024)

    2024

  56. [64]

    Uhl, J.H., Leyk, S., Chiang, Y.Y., Duan, W., Knoblock, C.A.: Extracting human settlement footprint from historical topographic map series using context-based machinelearning.In:8thInternationalConferenceofPatternRecognitionSystems. pp. 1–6 (2017)

    2017

  57. [65]

    IET Image Processing12(11), 2084–2091 (2018)

    Uhl, J.H., Leyk, S., Chiang, Y.Y., Duan, W., Knoblock, C.A.: Spatialising un- certainty in image segmentation using weakly supervised convolutional neural net- works: A case study from historical map processing. IET Image Processing12(11), 2084–2091 (2018)

    2018

  58. [66]

    Uhl, J.H., Leyk, S., Chiang, Y.Y., Duan, W., Knoblock, C.A.: Automated extrac- tion of human settlement patterns from historical topographic map series using weaklysupervisedconvolutionalneuralnetworks.IEEEAccess 8,6978–6996(2019)

    2019

  59. [67]

    Computers, En- vironment and Urban Systems94, 101794 (2022)

    Uhl, J.H., Leyk, S., Chiang, Y.Y., Knoblock, C.A.: Towards the automated large- scale reconstruction of past road networks from historical maps. Computers, En- vironment and Urban Systems94, 101794 (2022)

    2022

  60. [68]

    Remote Sensing13(18), 3672 (2021)

    Uhl, J.H., Leyk, S., Li, Z., Duan, W., Shbita, B., Chiang, Y.Y., Knoblock, C.A.: Combining remote-sensing-derived data and historical maps for long-term back- casting of urban extents. Remote Sensing13(18), 3672 (2021)

    2021

  61. [69]

    Remote Sensing12(24), 4048 (2020)

    Ulloa-Torrealba, Y., Stahlmann, R., Wegmann, M., Koellner, T.: Over 150 years of change: Object-oriented analysis of historical land cover in the main river catch- ment, Bavaria/Germany. Remote Sensing12(24), 4048 (2020)

    2020

  62. [70]

    Scientific Reports 12(1), 18997 (2022)

    Wong, C.S., Liao, H.M., Tsai, R.T.H., Chang, M.C.: Semi-supervised learning for topographic map analysis over time: A study of bridge segmentation. Scientific Reports 12(1), 18997 (2022)

    2022

  63. [71]

    GIScience & Remote Sensing59(1), 200–214 (2022)

    Wu, S., Heitzler, M., Hurni, L.: Leveraging uncertainty estimation and spatial pyramid pooling for extracting hydrological features from scanned historical topo- graphic maps. GIScience & Remote Sensing59(1), 200–214 (2022)

    2022

  64. [72]

    ISPRS Journal of Photogrammetry and Remote Sensing197, 199–211 (2023)

    Wu, S., Schindler, K., Heitzler, M., Hurni, L.: Domain adaptation in segment- ing historical maps: A weakly supervised approach through spatial co-occurrence. ISPRS Journal of Photogrammetry and Remote Sensing197, 199–211 (2023)

    2023

  65. [73]

    IEEE Access 8, 176704–176716 (2020)

    Zhang, Y., Yin, Y., Zimmermann, R., Wang, G., Varadarajan, J., Ng, S.K.: An en- hanced GAN model for automatic satellite-to-map image conversion. IEEE Access 8, 176704–176716 (2020)

    2020

  66. [74]

    ISPRS International Journal of Geo-Information 11(11), 572 (2022)

    Zhao, Y., Wang, G., Yang, J., Zhang, L., Qi, X.: Building block extraction from historical maps using deep object attention networks. ISPRS International Journal of Geo-Information 11(11), 572 (2022)

    2022

  67. [75]

    In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support

    Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: UNet++: A nested U-Net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. pp. 3–11 (2018)

    2018

  68. [76]

    In: Proceedings of the IEEE Interna- tional Conference on Computer Vision

    Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE Interna- tional Conference on Computer Vision. pp. 2223–2232 (2017)

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.